OLMo is a repository for training and using AI2's state-of-the-art open language models. It is built by scientists, for scientists.
First install PyTorch according to the instructions specific to your operating system.
To install from source (recommended for training/fine-tuning) run:
git clone https://github.com/allenai/OLMo.git
cd OLMo
pip install -e .[all]
Otherwise you can install the model code by itself directly from PyPI with:
pip install ai2-olmo
The core models in the OLMo family released so far are (all trained on the Dolma dataset):
Model | Training Tokens | Context Length | Training Config | W&B Logs | Data Order File(s) ☨ |
---|---|---|---|---|---|
OLMo 1B | 3 Trillion | 2048 | configs/official/OLMo-1B.yaml | wandb.ai/…/OLMo-1B | epoch 1 |
OLMo 7B | 2.5 Trillion | 2048 | configs/official/OLMo-7B.yaml | wandb.ai/…/OLMo-7B | epoch 1, epoch 2 |
OLMo 7B Twin 2T | 2 Trillion | 2048 | configs/official/OLMo-7B.yaml | wandb.ai/…/OLMo-7B-Twin-2T | epoch 1 |
OLMo 7B April 2024 | 2.05 Trillion | 4096 | configs/official/OLMo-7B-0424.yaml | Coming soon | Coming soon |
OLMo 7B July 2024 | 2.75 Trillion | 4096 | configs/official/OLMo-7B-0724.yaml | Coming soon | Coming soon |
☨ See Inspecting training data below for usage.
URLs to checkpoints at intermediate steps of the models' trainings can be found in the csv files under checkpoints/official/
. These 'directory' URLs cannot currently be directly accessed, but files within the directory are publicly accessible. These URLs can also be provided to the training script to resume training from the checkpoint (see Training). Each checkpoint directory consists of:
config.yaml
: the config at that training step.model.pt
,optim.pt
,train.pt
: model, optimizer and training state at that training step.
Details about the other types of OLMo checkpoints (including OLMo HF Transformers checkpoints) can be found in Checkpoints.md.
You can utilize our Hugging Face integration to run inference on the OLMo Transformers checkpoints:
from transformers import AutoModelForCausalLM, AutoTokenizer
olmo = AutoModelForCausalLM.from_pretrained("allenai/OLMo-7B-0724-hf")
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-7B-0724-hf")
message = ["Language modeling is "]
inputs = tokenizer(message, return_tensors='pt', return_token_type_ids=False)
response = olmo.generate(**inputs, max_new_tokens=100, do_sample=True, top_k=50, top_p=0.95)
print(tokenizer.batch_decode(response, skip_special_tokens=True)[0])
Alternatively, with the Hugging Face pipeline abstraction:
from transformers import pipeline
olmo_pipe = pipeline("text-generation", model="allenai/OLMo-7B-0724-hf")
print(olmo_pipe("Language modeling is"))
If you finetune the model using the code in Fine-tuning, you can use the conversion script to convert a native OLMo checkpoint to a Hugging Face-compatible checkpoint.
python scripts/convert_olmo_to_hf_new.py --input_dir /path/to/olmo/checkpoint --output_dir /path/to/hf/checkpoint/ --tokenizer_json_path tokenizers/allenai_gpt-neox-olmo-dolma-v1_5.json
olmo = AutoModelForCausalLM.from_pretrained("allenai/OLMo-7B-0724-hf", torch_dtype=torch.float16, load_in_8bit=True) # requires bitsandbytes
The quantized model is more sensitive to typing / cuda, so it is recommended to pass the inputs as inputs.input_ids.to('cuda') to avoid potential issues.
The configs used to train the official OLMo models are provided in the configs/official/
directory.
Note that while the training and validation data is public and free to download, the paths to the data within those configs are pointed at a CloudFlare R2 bucket, which requires an API key for programmatic access. So in order to use any of these configs to reproduce a training run you'll first have to download the corresponding data to a location of your choosing and then update the paths in the config accordingly.
You can derive the public HTTP URL from an R2 URL by replacing r2:https://olmo-data
with https://olmo-data.org
.
For example, if the R2 data URL is:
r2:https://olmo-data/preprocessed/olmo-mix/v1_5/gpt-neox-20b-pii-special/part-000-00000.npy
then the corresponding public URL is:
https://olmo-data.org/preprocessed/olmo-mix/v1_5/gpt-neox-20b-pii-special/part-000-00000.npy
Once you've updated the data paths in the config you can launch a training run via torchrun
. For example, to launch the 1B model training on a single 8x GPU node, you would run:
torchrun --nproc_per_node=8 scripts/train.py configs/official/OLMo-1B.yaml
You can use the same method to launch multi-node jobs as well. See the documentation for torchrun
to understand the additional arguments you'll need to configure the rendezvous backend / endpoint.
To resume training from a checkpoint, you can pass its path (local or URL)
to scripts/train.py
with the --load_path
arguments. For example, to resume training from step 1000 of the OLMo 1B run:
torchrun --nproc_per_node=8 scripts/train.py configs/official/OLMo-1B.yaml --load_path=https://olmo-checkpoints.org/ai2-llm/olmo-small/w1r5xfzt/step1000-unsharded
You may be interested in inspecting the exact tokens that composed a particular batch during the training of one of the OLMo models. We provide tools to do this, but first you'll need to download the data as above (unless you have an R2 API key) and update the corresponding config accordingly.
Then take note of the URL of the data order file you want, which can be found in the Models Overview table. For example, the data order file for the first epoch of the OLMo-7B model is https://olmo-checkpoints.org/ai2-llm/olmo-medium/wvc30anm/train_data/global_indices.npy.
Once you have that you can use this snippet to inspect the data within a particular batch:
import numpy as np
from cached_path import cached_path
from olmo.config import TrainConfig
from olmo.data import build_memmap_dataset
# Update these paths to what you want:
data_order_file_path = cached_path("https://olmo-checkpoints.org/ai2-llm/olmo-medium/wvc30anm/train_data/global_indices.npy")
train_config_path = "configs/official/OLMo-7B.yaml"
cfg = TrainConfig.load(train_config_path)
dataset = build_memmap_dataset(cfg, cfg.data)
batch_size = cfg.global_train_batch_size
global_indices = np.memmap(data_order_file_path, mode="r+", dtype=np.uint32)
def get_batch_instances(batch_idx: int) -> list[list[int]]:
batch_start = batch_idx * batch_size
batch_end = (batch_idx + 1) * batch_size
batch_indices = global_indices[batch_start:batch_end]
batch_instances = []
for index in batch_indices:
token_ids = dataset[index]["input_ids"].tolist()
batch_instances.append(token_ids)
return batch_instances
# Get all 2048 x 2048 token IDs in the first batch.
get_batch_instances(0)
To fine-tune an OLMo model using our trainer you'll first need to prepare your dataset by tokenizing it and saving the tokens IDs to a flat numpy memory-mapped array. See scripts/prepare_tulu_data.py
for an example with the Tulu V2 dataset, which can be easily modified for other datasets.
Next, prepare your training config. There are many examples in the configs/
directory that you can use as a starting point. The most important thing is to make sure the model parameters (the model
field in the config) match up with the checkpoint you're starting from. To be safe you can always start from the config that comes with the model checkpoint. At a minimum you'll need to make the following changes to the config or provide the corresponding overrides from the command line:
- Update
load_path
to point to the checkpoint you want to start from. - Set
reset_trainer_state
totrue
. - Update
data.paths
to point to thetoken_ids.npy
file you generated. - Optionally update
data.label_mask_paths
to point to thelabel_mask.npy
file you generated, unless you don't need special masking for the loss. - Update
evaluators
to add/remove in-loop evaluations.
Once you're satisfied with your training config, you can launch the training job via torchrun
. For example:
torchrun --nproc_per_node=8 scripts/train.py {path_to_train_config} \
--data.paths=[{path_to_data}/input_ids.npy] \
--data.label_mask_paths=[{path_to_data}/label_mask.npy] \
--load_path={path_to_checkpoint} \
--reset_trainer_state
Note: passing CLI overrides like --reset_trainer_state
is only necessary if you didn't update those fields in your config.
Additional tools for evaluating OLMo models are available at the OLMo Eval repo.
See Debugging.
@article{OLMo,
title={OLMo: Accelerating the Science of Language Models},
author={Dirk Groeneveld and Iz Beltagy and Pete Walsh and Akshita Bhagia and Rodney Kinney and Oyvind Tafjord and A. Jha and Hamish Ivison and Ian Magnusson and Yizhong Wang and Shane Arora and David Atkinson and Russell Authur and Khyathi Raghavi Chandu and Arman Cohan and Jennifer Dumas and Yanai Elazar and Yuling Gu and Jack Hessel and Tushar Khot and William Merrill and Jacob Daniel Morrison and Niklas Muennighoff and Aakanksha Naik and Crystal Nam and Matthew E. Peters and Valentina Pyatkin and Abhilasha Ravichander and Dustin Schwenk and Saurabh Shah and Will Smith and Emma Strubell and Nishant Subramani and Mitchell Wortsman and Pradeep Dasigi and Nathan Lambert and Kyle Richardson and Luke Zettlemoyer and Jesse Dodge and Kyle Lo and Luca Soldaini and Noah A. Smith and Hanna Hajishirzi},
year={2024},
url={https://api.semanticscholar.org/CorpusID:267365485},
journal={arXiv preprint},
}